Breast image recognition method based on non-enhanced nuclear magnetic resonance imaging
By fusing training data sets and network models with non-enhanced and enhanced NMR image features, the problem of contrast agent dependence in breast cancer detection is solved, and high-precision breast lesion recognition is achieved, saving time and cost.
Patent Information
- Application Number
- CN202510576254.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-06
- Publication Date
- 2025-07-22
AI Technical Summary
In the prior art, deep learning-based breast cancer detection methods mainly rely on contrast agent-enhanced MRI, which have problems with long imaging time, high cost and potential health risks. At the same time, the non-enhanced MRI image contrast is low, making it difficult to effectively identify subtle lesions.
By preprocessing breast images before and after injection of contrast agent, a training data set is constructed, and using learning networks and clinical diagnostic network models, the non-enhanced and enhanced NMR image features are fused to achieve breast images recognition.
Without the use of contrast agent, the breast lesion recognition accuracy comparable to that of enhanced NMR images is achieved using non-enhanced NMR images, saving imaging time and reducing contrast agent dependence.
Smart Images
Figure CN120355697A_ABST
Abstract
Description
Technical Field
[0001] This specification relates to the technical field of medical image classification, and particularly to a method for breast image recognition based on non-contrast magnetic resonance imaging. Background Art
[0002] Breast cancer is one of the most common cancers among female diseases, and its mortality rate is second only to lung cancer, seriously threatening women's health and quality of life. Therefore, early detection and treatment of breast cancer are of great significance. With the progress of medical imaging equipment, images containing accurate anatomical information can be acquired. Among them, magnetic resonance imaging (MRI) has unique advantages in breast examinations due to its high soft tissue resolution and radiation-free characteristics. Breast MRI examinations usually include multiple scanning sequences that can provide different imaging information. Depending on the scanning protocol, dozens to hundreds of images may be generated during a single breast cancer MRI examination.
[0003] Dynamic contrast-enhanced MRI (DCE-MRI) with the aid of a contrast agent, by injecting gadolinium-based contrast agents, helps to clearly show breast tumors and their degree of invasion of surrounding tissues. Therefore, existing deep learning-based image recognition methods mainly focus on DCE-MRI. However, the imaging time of DCE-MRI is relatively long, often occupying the main time of the entire breast MR examination. At the same time, gadolinium-based contrast agents are expensive and may cause problems such as in vivo deposition, allergy, and renal failure to patients, and need to be used after careful evaluation. Non-contrast enhanced MRI (NCE-MRI), such as T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), apparent diffusion coefficient (ADC), diffusion-weighted imaging (DWI), etc. sequences do not require injecting gadolinium-based contrast agents into patients, which can effectively solve the above problems. However, the image contrast of NCE-MRI is relatively low, making it difficult to show some subtle structures and lesions, and the recognition efficiency and accuracy are inferior to DCE-MRI. Summary of the Invention
[0004] Aiming at the above deficiencies in the prior art, a method for breast image recognition based on non-contrast magnetic resonance imaging provided by the present invention solves the problem of how to detect breast images only using non-contrast magnetic resonance images, and provides a reliable recognition solution for reducing the use of contrast agents and lowering the examination cost.
[0005] To achieve the above-mentioned invention purpose, the technical solution adopted by the present invention is as follows: A method for breast image recognition based on non-enhanced magnetic resonance imaging, including: S1: By preprocessing the breast images before and after injecting a contrast agent, a first training data set and a second training data set are obtained; the second training data set is a non-enhanced magnetic resonance image data set; S2: Using the first training data set, train the learning network to obtain the feature extraction layer parameters; S3: Input the second training data set and the feature extraction layer parameters into the clinical diagnosis network model, and through training, obtain a trained clinical diagnosis network model; S4: Using the trained clinical diagnosis network model, identify the non-enhanced magnetic resonance images to obtain the breast image recognition result, and complete the recognition of the breast images.
[0006] Further, the S1 includes: Collect breast images before and after injecting a contrast agent to obtain non-enhanced magnetic resonance images and enhanced magnetic resonance images respectively; Perform spatial registration on the non-enhanced magnetic resonance images and the enhanced magnetic resonance images to obtain registered image data; Perform data augmentation processing on the registered image data to obtain a breast image sample set; Divide the breast image sample set to obtain a first training data set and a second training data set.
[0007] Further, the S2 includes: Input the first training data set into the feature extraction layer of the learning network, and through feature extraction, obtain non-enhanced magnetic resonance features and enhanced magnetic resonance features respectively; Use the feature fusion layer of the learning network to fuse the non-enhanced magnetic resonance features and the enhanced magnetic resonance features to obtain fused features; Use the classifier layer of the learning network to identify the fused features to obtain the breast image classification result output by the learning network; Based on the breast image classification result output by the learning network, use the first loss function to train the learning network to obtain the feature extraction layer parameters.
[0008] Further, the expression of the first loss function is: ; ; ; Wherein, represents the result of the first loss function, represents the cross - entropy loss, represents the adjustment coefficient, represents the supplementary loss term, represents the number of samples in the dataset, represents the true label of the th sample, represents the non - enhanced MRI features, represents the enhanced MRI features, represents the L2 norm.
[0009] Furthermore, the expression of the parameters of the feature extraction layer is: ; wherein, represents the parameters of the feature extraction layer, represents the parameters of the feature extraction layer before update, represents the learning rate, represents the first - order momentum of the gradient, represents the second - order momentum of the gradient, represents a stability term to avoid division by zero, represents the weight decay coefficient.
[0010] Furthermore, the clinical diagnosis network model includes: A feature extraction layer, including a first feature extraction module and a second feature extraction module, which is used to extract features from the second training dataset to obtain a first feature vector, a second feature vector, and a third feature vector; A feature fusion layer, which is used to fuse the first feature vector and the second feature vector to obtain a fusion result; and splice the fusion result with the third feature vector to obtain a breast image feature vector; A classifier layer, which is used to analyze the breast image feature vector to obtain a breast image prediction result.
[0011] Furthermore, the S3 includes: Taking the parameters of the feature extraction layer as the parameters of the first feature extraction module, and using the first feature extraction module with fixed parameters to extract features from the second training dataset to obtain a first feature vector and a second feature vector; Inputting the second training dataset into the second feature extraction module, and through feature extraction, obtaining a third feature vector; Fusing the first feature vector and the second feature vector, splicing the obtained fusion result with the third feature vector to obtain a breast image feature vector; Analyzing the breast image feature vector to obtain a breast image prediction result; Based on the prediction result of the breast image, the clinical diagnosis network model is trained using the second loss function to optimize the parameters of the second feature extraction module, the feature fusion layer, and the classifier layer, and a trained clinical diagnosis network model is obtained.
[0012] The beneficial effects of the present invention are as follows: A method for breast image recognition based on non-enhanced magnetic resonance imaging. (1) In the recognition stage of breast lesions, only non-enhanced magnetic resonance images of the breast are required, and no contrast agent is needed to obtain them. In other related breast image classification methods, enhanced magnetic resonance images are required, saving magnetic resonance imaging time and reducing contrast agent dependence. (2) In the design of the clinical diagnosis network model, the present invention comprehensively considers the fused feature vectors representing non-enhanced magnetic resonance images and enhanced magnetic resonance images and the independent feature vectors only for non-enhanced magnetic resonance images. The former is obtained through the feature extractor after the pre-trained parameters obtained by the learning network in the previous training stage are transferred to the clinical diagnosis network model and frozen. This method ensures that in breast image recognition, even based only on low-contrast non-enhanced magnetic resonance images, breast image classification accuracy comparable to that of methods based on high-contrast enhanced magnetic resonance images can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] This specification will further illustrate in the form of exemplary embodiments, and these exemplary embodiments will be described in detail through the drawings. These embodiments are not restrictive. In these embodiments, the same numbers represent the same structures, where: Figure 1 is an exemplary flowchart of a method for breast image recognition based on non-enhanced magnetic resonance imaging according to some embodiments of this specification. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0014] The following describes the specific embodiments of the present invention to facilitate those skilled in the art of the present technology to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art of the present technology, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.
[0015] Embodiment Figure 1 is an exemplary flowchart of a method for breast image recognition based on non-enhanced magnetic resonance imaging according to some embodiments of this specification. As Figure 1 shown, the process includes the following steps. In some embodiments, the process can be executed by a processor.
[0016] S1: Preprocess the breast images before and after injecting a contrast agent to obtain a first training dataset and a second training dataset.
[0017] The breast images are the breast imaging data of patients who have undergone breast MRI examinations.
[0018] The first training dataset is an image dataset used to train a learning network. For example, the first training dataset can include a non-enhanced nuclear magnetic resonance image dataset and a corresponding enhanced nuclear magnetic resonance image dataset.
[0019] The second training dataset is an image dataset used to train a clinical diagnosis network model. For example, the second training dataset can include a non-enhanced nuclear magnetic resonance image dataset.
[0020] In some embodiments, the processor can obtain a test set based on the preprocessed breast images through partitioning.
[0021] The test set is a non-enhanced nuclear magnetic resonance image dataset used to test the performance of the clinical diagnosis network model.
[0022] In some embodiments, the processor can implement S1 based on the following steps: acquire breast lesion images before and after injecting a contrast agent to obtain non-enhanced nuclear magnetic resonance images and enhanced nuclear magnetic resonance images respectively; perform spatial registration on the non-enhanced nuclear magnetic resonance images and the enhanced nuclear magnetic resonance images to obtain registered image data; perform data augmentation processing on the registered image data to obtain a breast image sample set; partition the breast image sample set to obtain a first training dataset and a second training dataset.
[0023] Non-enhanced nuclear magnetic resonance images (NCE-MRI) are images acquired from patients before injecting a contrast agent. For example, nuclear magnetic resonance images can include image sequences such as T1WI, T2WI, DWI (diffusion factor b = 1000), and ADC images.
[0024] Enhanced nuclear magnetic resonance images (DCE-MRI) are images acquired from patients after injecting a contrast agent. For example, enhanced nuclear magnetic resonance images can include image sequences such as DCE2, DCE3, DCE4, and DCE5.
[0025] The registered image data is image data obtained by performing spatial registration on other image sequences using one image sequence as a standard.
[0026] In some embodiments, the processor can use T2WI as the standard and use ITK-SNAP software to register other sequences to obtain registered image data.
[0027] The breast image sample set is a training dataset obtained by performing data augmentation on benign breast lesion images.
[0028] In some embodiments, the processor may perform data augmentation processing such as flipping, translation, and rotation on the benign breast lesion samples in the training dataset to obtain a lesion sample set.
[0029] In some embodiments, the processor may randomly partition the lesion sample set according to non-enhanced magnetic resonance images to obtain a first training dataset (n = 666) with non-enhanced magnetic resonance images and corresponding enhanced magnetic resonance images, a second training dataset (n = 400) with non-enhanced magnetic resonance images, and a test set (n = 268).
[0030] S2: Use the first training dataset to train the learning network to obtain the feature extraction layer parameters.
[0031] The feature extraction layer parameters are the parameters of the feature extraction layer in the trained learning network.
[0032] In some embodiments, the processor may implement S2 based on the following steps: input the first training dataset into the feature extraction layer of the learning network, and through feature extraction, obtain non-enhanced magnetic resonance features and enhanced magnetic resonance features respectively; use the feature fusion layer of the learning network to fuse the non-enhanced magnetic resonance features and the enhanced magnetic resonance features to obtain fused features; use the classifier layer of the learning network to identify the fused features to obtain the breast image classification result output by the learning network; based on the breast image classification result output by the learning network, use the first loss function to train the learning network to obtain the feature extraction layer parameters.
[0033] The non-enhanced magnetic resonance features are features that reflect the lesion-related information in the non-enhanced magnetic resonance images.
[0034] The enhanced magnetic resonance features are features that reflect the lesion-related information in the enhanced magnetic resonance images.
[0035] The fused features are features that reflect the comprehensive information of the non-enhanced magnetic resonance features and the enhanced magnetic resonance features.
[0036] In some embodiments, the expression of the fused features may be: ; where represents the fused features, represents the non-enhanced magnetic resonance features, represents the enhanced magnetic resonance features.
[0037] The breast image classification result is the result of the breast lesion classification situation output by the learning network. For example, the breast image classification result may include benign lesions and malignant lesions, etc.
[0038] In some embodiments, the processor may calculate the error between the predicted value and the true label through a first loss function based on the breast image classification result output by the classifier layer, backpropagate to optimize the entire learning network, and update the parameters of the feature extraction layer and the classifier layer. After training, the pre-trained parameters of the feature extraction layer are obtained as the parameters of the feature extraction layer.
[0039] In some embodiments, the expression of the first loss function may be: ; ; ; where, represents the result of the first loss function, represents the cross-entropy loss, represents the adjustment coefficient, represents the supplementary loss term, represents the number of samples in the dataset, represents the true label of the th sample, represents the breast lesion classification result output by the learning network, represents the non-enhanced MRI feature, represents the enhanced MRI feature,
[0040] In some embodiments, the expression of the feature extraction layer parameters may be: ; where, represents the feature extraction layer parameters, represents the feature extraction layer parameters before update, represents the learning rate, represents the first-order momentum of the gradient, represents the second-order momentum of the gradient, represents the stability term to avoid division by zero, represents the weight decay coefficient.
[0041] In some embodiments, the processor may set the parameters, including: can be set to 0.0028, can be set to 1e-5, can be set to 1e-8.
[0042] S3: Input the second training dataset and the feature extraction layer parameters into the clinical diagnosis network model, and through training, obtain the trained clinical diagnosis network model.
[0043] The clinical diagnosis network model is used to identify breast lesion images and obtain the prediction results of breast lesions.
[0044] In some embodiments, the input of the clinical diagnosis network model can be the second training dataset, and the output of the clinical diagnosis network model can be the breast image prediction result.
[0045] In some embodiments, the structure of the clinical diagnosis network model is as follows: The clinical diagnosis network model (CD-Net) includes a feature extraction layer, a feature fusion layer, and a classifier layer. The output of the feature extraction layer serves as the input of the feature fusion layer, the output of the feature fusion layer serves as the input of the classifier layer, and the output of the classifier layer serves as the final output of the clinical diagnosis network model.
[0046] The feature extraction layer includes a first feature extraction module and a second feature extraction module, and is used to extract features from the non-enhanced magnetic resonance images in the second training dataset to obtain a first feature vector, a second feature vector, and a third feature vector. The input of the feature extraction layer can include the magnetic resonance images in the second training dataset, and the output can include the first feature vector, the second feature vector, and the third feature vector.
[0047] The first feature vector reflects the feature vector of the non-enhanced magnetic resonance image information in the second training dataset.
[0048] The second feature vector is a feature vector that infers the possible corresponding enhanced magnetic resonance image information based on the first feature vector.
[0049] The third feature vector is a feature vector that characterizes the non-enhanced magnetic resonance image information of actual clinical lesions.
[0050] The feature fusion layer is used to fuse the first feature vector and the second feature vector, splice the obtained fusion result with the third feature vector to obtain a breast image feature vector. The input of the feature fusion layer can include the first feature vector, the second feature vector, and the third feature vector, and the output can include the breast image feature vector.
[0051] The breast image feature vector is a feature vector that synthesizes the information of the first feature vector, the second feature vector, and the third feature vector.
[0052] The classifier layer is used to analyze the breast image feature vector to obtain a breast image prediction result. The input of the classifier layer can include the breast image feature vector, and the output can include the breast image prediction result.
[0053] The breast image prediction result is the prediction result of the breast lesion situation of the non-enhanced magnetic resonance image output by the clinical diagnosis network model.
[0054] In some embodiments, the processor may implement S4 based on the following steps: use the feature extraction layer parameters as the parameters of the first feature extraction module, and use the first feature extraction module with fixed parameters to extract features from the nuclear magnetic resonance images in the second training dataset to obtain a first feature vector and a second feature vector; input the nuclear magnetic resonance images in the second training dataset into the second feature extraction module, and through feature extraction, obtain a third feature vector; fuse the first feature vector and the second feature vector, splice the obtained fusion result with the third feature vector to obtain a breast image feature vector; analyze the breast image feature vector to obtain a breast image prediction result; based on the breast image prediction result, use the second loss function and the label to train the clinical diagnosis network model, and optimize the parameters of the second feature extraction module, the feature fusion layer, and the classifier layer to obtain a trained clinical diagnosis network model.
[0055] The second loss function is a loss function used to train the clinical diagnosis network model. For example, the second loss function may include a cross-entropy loss function, etc.
[0056] The label may be the actual situation of breast lesions corresponding to the image. The label can be manually marked.
[0057] The fusion result is a result reflecting the correlation between the breast nuclear magnetic resonance image and the enhanced nuclear magnetic resonance image of the same case.
[0058] In some embodiments, the expression of the fusion result may be: ; where represents the fusion result, represents the first feature vector, represents the second feature vector.
[0059] In some embodiments, the expression of the breast image feature vector may be: ; where represents the breast image feature vector, represents the third feature vector.
[0060] In some embodiments, the processor may set the initial learning rate of the clinical diagnosis network model to 0.0035, and the settings of other parameters are the same as those of the learning network (L-Net).
[0061] S4: Use the trained clinical diagnosis network model to identify the non-enhanced nuclear magnetic resonance image to obtain a breast image recognition result, and complete the recognition of the breast image.
[0062] The recognition result of breast images is the final recognition result of breast lesion images in non-enhanced magnetic resonance imaging. For example, the breast lesion recognition result may include benign lesions and malignant lesions, etc.
[0063] In some embodiments, the processor may use the training set to perform a performance test on the trained clinical diagnosis network model to obtain the performance of the trained clinical diagnosis network model; wherein, the specific performance of the trained clinical diagnosis network model is shown in Table 1.
[0064] Table 1 Performance Table of the Trained Clinical Diagnosis Network Model
[0065] In some embodiments of this specification, a breast image recognition method based on non-enhanced magnetic resonance imaging is proposed. (1) In the recognition stage of breast lesions, only non-enhanced magnetic resonance images of the breast are required and can be obtained without the need for contrast agents. In other related breast image classification methods, enhanced magnetic resonance images are required, saving magnetic resonance imaging time and reducing contrast agent dependence. (2) In the design of the clinical diagnosis network model, the present invention comprehensively considers the fused feature vectors representing non-enhanced magnetic resonance images and enhanced magnetic resonance images and the independent feature vectors only for non-enhanced magnetic resonance images. The former is obtained through the feature extractor after transferring the pre-trained parameters obtained by the learning network in the previous training stage to the clinical diagnosis network model and freezing them. This method ensures that in breast image recognition, even based only on low-contrast non-enhanced magnetic resonance images, a breast image classification accuracy comparable to that of the method based on high-contrast enhanced magnetic resonance images can be achieved.
Claims
1. A breast image recognition method based on non-enhanced magnetic resonance imaging, characterized in that, Including: S1: By preprocessing the breast images before and after injecting a contrast agent, a first training dataset and a second training dataset are obtained; The second training dataset is a non-enhanced magnetic resonance image dataset; S2: Using the first training dataset, the learning network is trained to obtain feature extraction layer parameters; S3: Inputting the second training dataset and the feature extraction layer parameters into the clinical diagnosis network model, and through training, a trained clinical diagnosis network model is obtained; S4: Using the trained clinical diagnosis network model, the non-enhanced magnetic resonance images are identified to obtain breast image recognition results, and the recognition of breast images is completed.
2. The method for breast image recognition based on non-enhanced nuclear magnetic resonance imaging according to claim 1, wherein The S1 includes: Collecting breast images before and after injecting a contrast agent to obtain non-enhanced magnetic resonance images and enhanced magnetic resonance images respectively; Performing spatial registration on the non-enhanced magnetic resonance images and the enhanced magnetic resonance images to obtain registered image data; Performing data augmentation processing on the registered image data to obtain a breast image sample set; Dividing the breast image sample set to obtain a first training dataset and a second training dataset.
3. The method for breast image recognition based on non-enhanced nuclear magnetic resonance imaging according to claim 1, wherein The S2 includes: Inputting the first training dataset into the feature extraction layer of the learning network, and through feature extraction, non-enhanced magnetic resonance features and enhanced magnetic resonance features are obtained respectively; Using the feature fusion layer of the learning network to fuse the non-enhanced magnetic resonance features and the enhanced magnetic resonance features to obtain fused features; Using the classifier layer of the learning network to identify the fused features to obtain the breast image classification result output by the learning network; Based on the breast image classification result output by the learning network, the learning network is trained using the first loss function to obtain feature extraction layer parameters.
4. The method for breast image recognition based on non-enhanced nuclear magnetic resonance imaging according to claim 3, wherein The expression of the first loss function is: ； ； ； Among them, represents the result of the first loss function, represents the cross-entropy loss, represents the adjustment coefficient, represents the supplementary loss term, represents the number of samples in the dataset, represents the true label of the th sample, represents the classification result of the breast image output by the learning network, represents the non-enhanced MRI features, represents the L2 norm.
5. The method for breast image recognition based on non-enhanced nuclear magnetic resonance imaging according to claim 3, wherein The expression of the feature extraction layer parameters is: ； Among them, represents the parameters of the updated feature extraction layer, represents the parameters of the feature extraction layer before update, represents the learning rate, represents the first-order momentum of the gradient, represents the second-order momentum of the gradient, represents a stabilization term to avoid division by zero, represents the weight decay coefficient.
6. The method for breast image recognition based on non-enhanced magnetic resonance imaging according to claim 2, wherein The clinical diagnosis network model includes: A feature extraction layer, including a first feature extraction module and a second feature extraction module, for extracting features from the second training dataset to obtain a first feature vector, a second feature vector, and a third feature vector; A feature fusion layer, for fusing the first feature vector and the second feature vector to obtain a fusion result; splicing the fusion result with the third feature vector to obtain a breast image feature vector; A classifier layer, for analyzing the breast image feature vector to obtain a breast image prediction result.
7. The method for breast image recognition based on non-enhanced nuclear magnetic resonance imaging according to claim 6, wherein The S3 includes: Taking the feature extraction layer parameters as the parameters of the first feature extraction module, and using the first feature extraction module with fixed parameters to extract features from the second training dataset to obtain a first feature vector and a second feature vector; Inputting the second training dataset into the second feature extraction module, and through feature extraction, a third feature vector is obtained; Fusing the first feature vector and the second feature vector, and splicing the obtained fusion result with the third feature vector to obtain a breast image feature vector; Analyzing the breast image feature vector to obtain a breast image prediction result; Based on the prediction result of the breast image, the clinical diagnosis network model is trained using the second loss function to optimize the parameters of the second feature extraction module, the feature fusion layer, and the classifier layer, and a trained clinical diagnosis network model is obtained.